Wasserstein Distance-Based Graph Kernel for Enhancing Drug Safety and Efficacy Prediction *
Mohammed Aburidi, Roummel F. Marcia · 2024
Drug development relies heavily on the optimization of therapeutic agents concerning their pharmacodynamics, pharmacokinetics, and toxicological properties. These properties, collectively known as ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity), play a pivotal role in determining a drug's efficacy and safety. However, assessing these properties during the early stages of drug development poses substantial challenges, driven by the resource-intensive nature of experimental evaluation and the scarcity of comprehensive data. To address these hurdles, computational and predictive tools have gained prominence, capitalizing on recent advances in machine learning and graph-based methodologies. This study introduces an innovative approach that leverages optimal transport (OT) to construct a graph kernel for predicting drug ADMET properties. The approach involves graph matching to create a similarity matrix, which is subsequently employed in a predictive model. Extensive evaluations on 19 ADMET datasets demonstrate the promise of this methodology. Our OT-based graph kernel out-performs state-of-the-art graph deep learning models in 8 out of 19 datasets, outperforming the most impactful GNN (which outperforms on 4 datasets) and competes effectively in 3 others, showcasing its adaptability and generalization capabilities and potential to enhance pharmaceuticals.